10 files changed, 502 insertions, 0 deletions
diff --git a/app/lstm_chem/__init__.py b/app/lstm_chem/__init__.py
new file mode 100755
index 0000000..8b13789
--- /dev/null
+++ b/app/lstm_chem/__init__.py
@@ -0,0 +1 @@
+
diff --git a/app/lstm_chem/data_loader.py b/app/lstm_chem/data_loader.py
new file mode 100755
index 0000000..86ddbba
--- /dev/null
+++ b/app/lstm_chem/data_loader.py
@@ -0,0 +1,122 @@
+import json
+import os
+import numpy as np
+from tqdm import tqdm
+from tensorflow.keras.utils import Sequence
+from lstm_chem.utils.smiles_tokenizer import SmilesTokenizer
+
+
+class DataLoader(Sequence):
+    def __init__(self, config, data_type='train'):
+        self.config = config
+        self.data_type = data_type
+        assert self.data_type in ['train', 'valid', 'finetune']
+
+        self.max_len = 0
+
+        if self.data_type == 'train':
+            self.smiles = self._load(self.config.data_filename)
+        elif self.data_type == 'finetune':
+            self.smiles = self._load(self.config.finetune_data_filename)
+        else:
+            pass
+
+        self.st = SmilesTokenizer()
+        self.one_hot_dict = self.st.one_hot_dict
+
+        self.tokenized_smiles = self._tokenize(self.smiles)
+
+        if self.data_type in ['train', 'valid']:
+            self.idx = np.arange(len(self.tokenized_smiles))
+            self.valid_size = int(
+                np.ceil(
+                    len(self.tokenized_smiles) * self.config.validation_split))
+            np.random.seed(self.config.seed)
+            np.random.shuffle(self.idx)
+
+    def _set_data(self):
+        if self.data_type == 'train':
+            ret = [
+                self.tokenized_smiles[self.idx[i]]
+                for i in self.idx[self.valid_size:]
+            ]
+        elif self.data_type == 'valid':
+            ret = [
+                self.tokenized_smiles[self.idx[i]]
+                for i in self.idx[:self.valid_size]
+            ]
+        else:
+            ret = self.tokenized_smiles
+        return ret
+
+    def _load(self, data_filename):
+        length = self.config.data_length
+        print('loading SMILES...')
+        with open(data_filename) as f:
+            smiles = [s.rstrip() for s in f]
+        if length != 0:
+            smiles = smiles[:length]
+        print('done.')
+        return smiles
+
+    def _tokenize(self, smiles):
+        assert isinstance(smiles, list)
+        print('tokenizing SMILES...')
+        tokenized_smiles = [self.st.tokenize(smi) for smi in tqdm(smiles)]
+
+        if self.data_type == 'train':
+            for tokenized_smi in tokenized_smiles:
+                length = len(tokenized_smi)
+                if self.max_len < length:
+                    self.max_len = length
+            self.config.train_smi_max_len = self.max_len
+        print('done.')
+        return tokenized_smiles
+
+    def __len__(self):
+        target_tokenized_smiles = self._set_data()
+        if self.data_type in ['train', 'valid']:
+            ret = int(
+                np.ceil(
+                    len(target_tokenized_smiles) /
+                    float(self.config.batch_size)))
+        else:
+            ret = int(
+                np.ceil(
+                    len(target_tokenized_smiles) /
+                    float(self.config.finetune_batch_size)))
+        return ret
+
+    def __getitem__(self, idx):
+        target_tokenized_smiles = self._set_data()
+        if self.data_type in ['train', 'valid']:
+            data = target_tokenized_smiles[idx *
+                                           self.config.batch_size:(idx + 1) *
+                                           self.config.batch_size]
+        else:
+            data = target_tokenized_smiles[idx *
+                                           self.config.finetune_batch_size:
+                                           (idx + 1) *
+                                           self.config.finetune_batch_size]
+        data = self._padding(data)
+
+        self.X, self.y = [], []
+        for tp_smi in data:
+            X = [self.one_hot_dict[symbol] for symbol in tp_smi[:-1]]
+            self.X.append(X)
+            y = [self.one_hot_dict[symbol] for symbol in tp_smi[1:]]
+            self.y.append(y)
+
+        self.X = np.array(self.X, dtype=np.float32)
+        self.y = np.array(self.y, dtype=np.float32)
+
+        return self.X, self.y
+
+    def _pad(self, tokenized_smi):
+        return ['G'] + tokenized_smi + ['E'] + [
+            'A' for _ in range(self.max_len - len(tokenized_smi))
+        ]
+
+    def _padding(self, data):
+        padded_smiles = [self._pad(t_smi) for t_smi in data]
+        return padded_smiles
diff --git a/app/lstm_chem/finetuner.py b/app/lstm_chem/finetuner.py
new file mode 100755
index 0000000..904958b
--- /dev/null
+++ b/app/lstm_chem/finetuner.py
@@ -0,0 +1,24 @@
+from lstm_chem.utils.smiles_tokenizer import SmilesTokenizer
+from lstm_chem.generator import LSTMChemGenerator
+
+
+class LSTMChemFinetuner(LSTMChemGenerator):
+    def __init__(self, modeler, finetune_data_loader):
+        self.session = modeler.session
+        self.model = modeler.model
+        self.config = modeler.config
+        self.finetune_data_loader = finetune_data_loader
+        self.st = SmilesTokenizer()
+
+    def finetune(self):
+        self.model.compile(optimizer=self.config.optimizer,
+                           loss='categorical_crossentropy')
+
+        history = self.model.fit_generator(
+            self.finetune_data_loader,
+            steps_per_epoch=self.finetune_data_loader.__len__(),
+            epochs=self.config.finetune_epochs,
+            verbose=self.config.verbose_training,
+            use_multiprocessing=True,
+            shuffle=True)
+        return history
diff --git a/app/lstm_chem/generator.py b/app/lstm_chem/generator.py
new file mode 100755
index 0000000..498f864
--- /dev/null
+++ b/app/lstm_chem/generator.py
@@ -0,0 +1,44 @@
+from tqdm import tqdm
+import numpy as np
+from lstm_chem.utils.smiles_tokenizer import SmilesTokenizer
+
+
+class LSTMChemGenerator(object):
+    def __init__(self, modeler):
+        self.session = modeler.session
+        self.model = modeler.model
+        self.config = modeler.config
+        self.st = SmilesTokenizer()
+
+    def _generate(self, sequence):
+        while (sequence[-1] != 'E') and (len(self.st.tokenize(sequence)) <=
+                                         self.config.smiles_max_length):
+            x = self.st.one_hot_encode(self.st.tokenize(sequence))
+            preds = self.model.predict_on_batch(x)[0][-1]
+            next_idx = self.sample_with_temp(preds)
+            sequence += self.st.table[next_idx]
+
+        sequence = sequence[1:].rstrip('E')
+        return sequence
+
+    def sample_with_temp(self, preds):
+        streched = np.log(preds) / self.config.sampling_temp
+        streched_probs = np.exp(streched) / np.sum(np.exp(streched))
+        return np.random.choice(range(len(streched)), p=streched_probs)
+
+    def sample(self, num=1, start='G'):
+        sampled = []
+        if self.session == 'generate':
+            for _ in tqdm(range(num)):
+                sampled.append(self._generate(start))
+            return sampled
+        else:
+            from rdkit import Chem, RDLogger
+            RDLogger.DisableLog('rdApp.*')
+            while len(sampled) < num:
+                sequence = self._generate(start)
+                mol = Chem.MolFromSmiles(sequence)
+                if mol is not None:
+                    canon_smiles = Chem.MolToSmiles(mol)
+                    sampled.append(canon_smiles)
+            return sampled
diff --git a/app/lstm_chem/model.py b/app/lstm_chem/model.py
new file mode 100755
index 0000000..079589a
--- /dev/null
+++ b/app/lstm_chem/model.py
@@ -0,0 +1,73 @@
+import os
+import time
+from tensorflow.keras import Sequential
+from tensorflow.keras.models import model_from_json
+from tensorflow.keras.layers import LSTM, Dense
+from tensorflow.keras.initializers import RandomNormal
+from lstm_chem.utils.smiles_tokenizer import SmilesTokenizer
+
+
+class LSTMChem(object):
+    def __init__(self, config, session='train'):
+        assert session in ['train', 'generate', 'finetune'], \
+                'one of {train, generate, finetune}'
+
+        self.config = config
+        self.session = session
+        self.model = None
+
+        if self.session == 'train':
+            self.build_model()
+        else:
+            self.model = self.load(self.config.model_arch_filename,
+                                   self.config.model_weight_filename)
+
+    def build_model(self):
+        st = SmilesTokenizer()
+        n_table = len(st.table)
+        weight_init = RandomNormal(mean=0.0,
+                                   stddev=0.05,
+                                   seed=self.config.seed)
+
+        self.model = Sequential()
+        self.model.add(
+            LSTM(units=self.config.units,
+                 input_shape=(None, n_table),
+                 return_sequences=True,
+                 kernel_initializer=weight_init,
+                 dropout=0.3))
+        self.model.add(
+            LSTM(units=self.config.units,
+                 input_shape=(None, n_table),
+                 return_sequences=True,
+                 kernel_initializer=weight_init,
+                 dropout=0.5))
+        self.model.add(
+            Dense(units=n_table,
+                  activation='softmax',
+                  kernel_initializer=weight_init))
+
+        arch = self.model.to_json(indent=2)
+        self.config.model_arch_filename = os.path.join(self.config.exp_dir,
+                                                       'model_arch.json')
+        with open(self.config.model_arch_filename, 'w') as f:
+            f.write(arch)
+
+        self.model.compile(optimizer=self.config.optimizer,
+                           loss='categorical_crossentropy')
+
+    def save(self, checkpoint_path):
+        assert self.model, 'You have to build the model first.'
+
+        print('Saving model ...')
+        self.model.save_weights(checkpoint_path)
+        print('model saved.')
+
+    def load(self, model_arch_file, checkpoint_file):
+        print(f'Loading model architecture from {model_arch_file} ...')
+        with open(model_arch_file) as f:
+            model = model_from_json(f.read())
+        print(f'Loading model checkpoint from {checkpoint_file} ...')
+        model.load_weights(checkpoint_file)
+        print('Loaded the Model.')
+        return model
diff --git a/app/lstm_chem/trainer.py b/app/lstm_chem/trainer.py
new file mode 100755
index 0000000..4e8057e
--- /dev/null
+++ b/app/lstm_chem/trainer.py
@@ -0,0 +1,56 @@
+from glob import glob
+import os
+from tensorflow.keras.callbacks import ModelCheckpoint, TensorBoard
+
+
+class LSTMChemTrainer(object):
+    def __init__(self, modeler, train_data_loader, valid_data_loader):
+        self.model = modeler.model
+        self.config = modeler.config
+        self.train_data_loader = train_data_loader
+        self.valid_data_loader = valid_data_loader
+        self.callbacks = []
+        self.init_callbacks()
+
+    def init_callbacks(self):
+        self.callbacks.append(
+            ModelCheckpoint(
+                filepath=os.path.join(
+                    self.config.checkpoint_dir,
+                    '%s-{epoch:02d}-{val_loss:.2f}.hdf5' %
+                    self.config.exp_name),
+                monitor=self.config.checkpoint_monitor,
+                mode=self.config.checkpoint_mode,
+                save_best_only=self.config.checkpoint_save_best_only,
+                save_weights_only=self.config.checkpoint_save_weights_only,
+                verbose=self.config.checkpoint_verbose,
+            ))
+        self.callbacks.append(
+            TensorBoard(
+                log_dir=self.config.tensorboard_log_dir,
+                write_graph=self.config.tensorboard_write_graph,
+            ))
+
+    def train(self):
+        history = self.model.fit_generator(
+            self.train_data_loader,
+            steps_per_epoch=self.train_data_loader.__len__(),
+            epochs=self.config.num_epochs,
+            verbose=self.config.verbose_training,
+            validation_data=self.valid_data_loader,
+            validation_steps=self.valid_data_loader.__len__(),
+            use_multiprocessing=True,
+            shuffle=True,
+            callbacks=self.callbacks)
+
+        last_weight_file = glob(
+            os.path.join(
+                f'{self.config.checkpoint_dir}',
+                f'{self.config.exp_name}-{self.config.num_epochs:02}*.hdf5')
+        )[0]
+
+        assert os.path.exists(last_weight_file)
+        self.config.model_weight_filename = last_weight_file
+
+        with open(os.path.join(self.config.exp_dir, 'config.json'), 'w') as f:
+            f.write(self.config.toJSON(indent=2))
diff --git a/app/lstm_chem/utils/config.py b/app/lstm_chem/utils/config.py
new file mode 100755
index 0000000..fff7359
--- /dev/null
+++ b/app/lstm_chem/utils/config.py
@@ -0,0 +1,26 @@
+import os
+import time
+import json
+from bunch import Bunch
+
+
+def get_config_from_json(json_file):
+    with open(json_file, 'r') as config_file:
+        config_dict = json.load(config_file)
+    config = Bunch(config_dict)
+    return config
+
+
+def process_config(json_file):
+    config = get_config_from_json(json_file)
+    config.config_file = json_file
+    config.exp_dir = os.path.join(
+        'experiments', time.strftime('%Y-%m-%d/', time.localtime()),
+        config.exp_name)
+    config.tensorboard_log_dir = os.path.join(
+        'experiments', time.strftime('%Y-%m-%d/', time.localtime()),
+        config.exp_name, 'logs/')
+    config.checkpoint_dir = os.path.join(
+        'experiments', time.strftime('%Y-%m-%d/', time.localtime()),
+        config.exp_name, 'checkpoints/')
+    return config
diff --git a/app/lstm_chem/utils/dirs.py b/app/lstm_chem/utils/dirs.py
new file mode 100755
index 0000000..bcd2a49
--- /dev/null
+++ b/app/lstm_chem/utils/dirs.py
@@ -0,0 +1,12 @@
+import os
+import sys
+
+
+def create_dirs(dirs):
+    try:
+        for dir_ in dirs:
+            if not os.path.exists(dir_):
+                os.makedirs(dir_)
+    except Exception as err:
+        print(f'Creating directories error: {err}')
+        sys.exit()
diff --git a/app/lstm_chem/utils/smiles_tokenizer.py b/app/lstm_chem/utils/smiles_tokenizer.py
new file mode 100755
index 0000000..d15d625
--- /dev/null
+++ b/app/lstm_chem/utils/smiles_tokenizer.py
@@ -0,0 +1,72 @@
+import copy
+import numpy as np
+
+import time
+
+
+class SmilesTokenizer(object):
+    def __init__(self):
+        atoms = [
+            'Li',
+            'Na',
+            'Al',
+            'Si',
+            'Cl',
+            'Sc',
+            'Zn',
+            'As',
+            'Se',
+            'Br',
+            'Sn',
+            'Te',
+            'Cn',
+            'H',
+            'B',
+            'C',
+            'N',
+            'O',
+            'F',
+            'P',
+            'S',
+            'K',
+            'V',
+            'I',
+        ]
+        special = [
+            '(', ')', '[', ']', '=', '#', '%', '0', '1', '2', '3', '4', '5',
+            '6', '7', '8', '9', '+', '-', 'se', 'te', 'c', 'n', 'o', 's'
+        ]
+        padding = ['G', 'A', 'E']
+
+        self.table = sorted(atoms, key=len, reverse=True) + special + padding
+        self.table_len = len(self.table)
+
+        self.one_hot_dict = {}
+        for i, symbol in enumerate(self.table):
+            vec = np.zeros(self.table_len, dtype=np.float32)
+            vec[i] = 1
+            self.one_hot_dict[symbol] = vec
+
+    def tokenize(self, smiles):
+        N = len(smiles)
+        i = 0
+        token = []
+
+        timeout = time.time() + 5   # 5 seconds from now
+        while (i < N):
+            for j in range(self.table_len):
+                symbol = self.table[j]
+                if symbol == smiles[i:i + len(symbol)]:
+                    token.append(symbol)
+                    i += len(symbol)
+                    break
+            if time.time() > timeout:
+                break 
+        return token
+
+    def one_hot_encode(self, tokenized_smiles):
+        result = np.array(
+            [self.one_hot_dict[symbol] for symbol in tokenized_smiles],
+            dtype=np.float32)
+        result = result.reshape(1, result.shape[0], result.shape[1])
+        return result
diff --git a/app/lstm_chem/utils/smiles_tokenizer.py.4faeffb638548d04ca4415dfe32cf8c7.py b/app/lstm_chem/utils/smiles_tokenizer.py.4faeffb638548d04ca4415dfe32cf8c7.py
new file mode 100755
index 0000000..d15d625
--- /dev/null
+++ b/app/lstm_chem/utils/smiles_tokenizer.py.4faeffb638548d04ca4415dfe32cf8c7.py
@@ -0,0 +1,72 @@
+import copy
+import numpy as np
+
+import time
+
+
+class SmilesTokenizer(object):
+    def __init__(self):
+        atoms = [
+            'Li',
+            'Na',
+            'Al',
+            'Si',
+            'Cl',
+            'Sc',
+            'Zn',
+            'As',
+            'Se',
+            'Br',
+            'Sn',
+            'Te',
+            'Cn',
+            'H',
+            'B',
+            'C',
+            'N',
+            'O',
+            'F',
+            'P',
+            'S',
+            'K',
+            'V',
+            'I',
+        ]
+        special = [
+            '(', ')', '[', ']', '=', '#', '%', '0', '1', '2', '3', '4', '5',
+            '6', '7', '8', '9', '+', '-', 'se', 'te', 'c', 'n', 'o', 's'
+        ]
+        padding = ['G', 'A', 'E']
+
+        self.table = sorted(atoms, key=len, reverse=True) + special + padding
+        self.table_len = len(self.table)
+
+        self.one_hot_dict = {}
+        for i, symbol in enumerate(self.table):
+            vec = np.zeros(self.table_len, dtype=np.float32)
+            vec[i] = 1
+            self.one_hot_dict[symbol] = vec
+
+    def tokenize(self, smiles):
+        N = len(smiles)
+        i = 0
+        token = []
+
+        timeout = time.time() + 5   # 5 seconds from now
+        while (i < N):
+            for j in range(self.table_len):
+                symbol = self.table[j]
+                if symbol == smiles[i:i + len(symbol)]:
+                    token.append(symbol)
+                    i += len(symbol)
+                    break
+            if time.time() > timeout:
+                break 
+        return token
+
+    def one_hot_encode(self, tokenized_smiles):
+        result = np.array(
+            [self.one_hot_dict[symbol] for symbol in tokenized_smiles],
+            dtype=np.float32)
+        result = result.reshape(1, result.shape[0], result.shape[1])
+        return result